The UK Just Broke Google's AI Search Bargain With Publishers

For years, Google search bundled two different uses of publisher content. The content could be crawled for the index, and it could also feed AI features layered above the results. A publisher that objected to the second use had no practical way to preserve the first. The cost of refusing was disappearance from the discovery channel that paid the bills.

On June 3, 2026, the UK's Competition and Markets Authority imposed a conduct requirement on Google search. The requirement sits inside Google's strategic market status designation under the UK's digital markets regime. It is not a court judgment and it is not a promise from Google. It is a standing rule with an implementation deadline.

The mechanism matters. In October 2025, the CMA designated Google with strategic market status in general search, finding substantial and entrenched market power. Four proposed conduct requirements went to consultation on January 28, 2026. The publisher requirement was then finalized and imposed.

The rule has three practical parts. Publishers must get effective tools to prevent their content from powering AI features such as AI Overviews without losing traditional search rankings. After consultation feedback, the scope also covers the use of publisher content for fine-tuning Google's AI models. And when publisher content does appear in an AI-generated result, the result must provide clear attribution and links.

Google has nine months to implement the full requirement, with key controls expected sooner, and must publish compliance reports every six months for the first year. The CMA is also watching Google's May 2026 AI-heavy search changes and has signaled further work on a fair exchange of value. The rule is narrow, but its consequences are not.

The old arrangement looked like a choice but functioned like a threat. Participate in Google's system on Google's terms, including whatever AI use Google attached to crawling, or give up the ranking visibility that made publishing commercially possible. The decision was “opt in” only in the technical sense.

That bundle produced a predictable result. Publishers accepted diffuse, long-term erosion in exchange for avoiding the immediate and measurable loss of search traffic. Google received the content, the training value, and the ability to summarize the work for a user who might never click. The publisher received a ranking position and the hope that a click would survive the answer layer.

The CMA's insight is not that AI features are inherently illegal. It is that the market power came from coupling separate products. A publisher could not accept the search index while declining the AI feature, so the apparent consent was not a meaningful negotiation.

This connects to the broader AI visibility problem. Being mentioned in an answer is not the same as receiving a visit, and being used as a source is not the same as being paid for the value of the source. The CMA rule addresses the control surface first. It does not settle the economics.

Unbundling is a bargaining change. A publisher that can credibly withhold content from AI features without committing search suicide has something Google must respond to. The requirement gives publishers leverage, not a guaranteed favorable outcome.

There is a reason to resist a victory lap. The requirement redistributes the dilemma rather than dissolving it. AI answers are becoming a place where people make decisions. A publisher that opts out protects a source asset, but it may also remove itself from the interface where a new audience is forming. The strategic cost of absence is rising.

The answer is not a company-wide switch. Segment the content by where its value lives. Commodity explainers and broad awareness content often exist to be found, so inclusion with clear attribution may be sensible. Original research, proprietary datasets, investigations, and content monetized through direct reading lose more when they are compressed into an answer. Those are stronger candidates for an opt-out or a licensing conversation.

The middle band is where the operating judgment sits. A product comparison may need search reach but also contain original evaluation. A local guide may benefit from AI discovery but lose its commercial value if the answer omits the business that produced it. Content policy should give owners a method for deciding, not a universal answer.

Attribution makes the opt-in side more measurable. If AI-generated results link clearly, track the referral, conversion quality, and assisted demand. If inclusion produces citations but no useful visits or brand recognition, that is evidence for revising the policy. The control is valuable because it creates a test, not because it dictates the result.

Put the decision in writing. State which classes stay in AI features, which classes opt out, which exceptions require legal or commercial review, and who revisits the rule as the answer layer changes. The default should be intentional, not inherited from a crawler.

The conduct requirement is UK-specific. It does not automatically change how Google treats publishers in the United States, the European Union, or other markets. That boundary should be kept clear because regulatory enthusiasm often gets reported as if a local rule were already global.

The model is still significant. The European Union has a gatekeeper regime under the Digital Markets Act and a publisher grievance pipeline. The Munich AI Overviews liability decision follows a different legal track, asking what Google-generated claims mean for responsibility, but it points at the same underlying reality: an AI search layer is not a neutral index.

Market-power regulation and content-liability cases approach the system from opposite directions. One asks whether the platform can bundle uses and force a default. The other asks whether the platform is responsible for what its generated answer says. Together they make the AI answer layer more governable than the old “just search” framing suggests.

For global publishers, the UK rollout is a live experiment. UK operations will have controls, compliance reports, and a bargaining position that other regions may not yet have. Create one global content-value inventory now, then attach the local rule set as each jurisdiction changes.

Do not predict the exact next regulation. Build the capability that will make the next rule easy to absorb: an inventory of source assets, an AI-use policy, a measurement baseline, and a process for evaluating whether the content's commercial value is discovery, direct readership, licensing, or some combination.

Publishers and content-led brands do not need to wait for Google's final interface before preparing. Five practical moves turn the CMA rule into an operating capability.

The CMA did not end the publisher-platform conflict. It made one part of the conflict negotiable. Search visibility and AI use can now be discussed as distinct choices, with a record of which choice was made and why.

That is a modest change in interface terms and a large change in power terms. Whatever a publisher decides, it will be a decision instead of a default.

A policy is not a control until someone can test it. The first test is separation. Ask whether a publisher can keep ordinary indexing active while refusing AI feature use. The second is scope. Ask whether the refusal reaches model fine-tuning as well as the answer surface. The third is evidence. Ask what Google reports, how often it reports it, and whether a publisher can compare that report with its own crawl and citation observations.

These are simple questions because the rule is trying to make a complex relationship legible. A publisher should not have to infer from traffic loss whether an opt-out affected ranking. It should be possible to observe the switch, the date it changed, the surfaces it affected, and the traditional ranking behavior that remained outside the switch.

Attribution needs the same discipline. “Source” is not enough if the link is buried, broken, or attached to a sentence the publisher did not write. Record the query, the generated answer, the visible source treatment, the landing page, and the referral. An attribution obligation creates a usable signal only when the publisher can inspect it at the level of an actual answer.

The nine-month clock should therefore be treated as a test window. Select a representative set of pages across the value classes. Choose queries that matter commercially and queries that surface original research. Record the current result, the current link, and the current visit pattern. When Google implements the controls, test the same set again.

The result will not be one universal answer. A publisher may discover that awareness pages benefit from inclusion while proprietary research performs better behind an opt-out. Another may find that clear attribution is worth more than the licensing value of withholding. The important thing is that the decision comes from the economics of the content, not from fear of turning off the wrong machine.

That is the practical meaning of the conduct requirement. It gives publishers a control surface. It does not choose the strategy for them.

The most defensible publisher strategy starts with an asset-level question: what does this page do for the business? A research report may earn subscriptions, citations, leads, licensing conversations, or authority for a product. A glossary may earn discovery and reduce support demand. A local guide may earn visits that turn into calls. Those values should not be hidden behind one domain-wide AI setting.

For each class, record the current path from exposure to outcome. If an AI answer cites the page, does the visitor arrive, read, subscribe, request a quote, or disappear? If a page is excluded, does search traffic remain stable while AI visibility changes? The CMA's separation makes these questions observable, but only if publishers collect a baseline before they flip anything.

Use the evidence to negotiate internally. Editorial may care about reach, commercial teams about qualified demand, and legal about control over original work. A shared inventory gives each team the same object to discuss. It also prevents the debate from collapsing into “more visibility is always good” or “all training use is theft.” Both statements are too broad to guide a portfolio.

The control is most valuable when it leads to a reversible decision. Start with a small, representative class, monitor the result, and record the reason for any change. That creates institutional memory as the answer layer evolves and as other regulators introduce their own versions of the choice.

The easiest mistake is to treat the new control as a binary moral test. Opting in does not mean a publisher approves every future use. Opting out does not mean the content has no discovery value. Both settings are tools in a portfolio, and both need review when the platform changes how answers are formed.

Keep the policy close to the evidence. When a page changes audience, revenue model, or source value, reclassify it. When Google's attribution changes, recheck the referral. When a new jurisdiction adds a control, map the local setting to the same global inventory. The work is repetitive, but it is considerably more useful than arguing about AI search in the abstract.